Set to Ordered Text: Generating Discharge Instructions from Medical Billing Codes

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 6165–6175, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 6165 Set to Ordered Text: Generating Discharge Instructions from Medical Billing Codes Litton J Kurisinkel, Nancy F. Chen Institute for Infocomm Research, A*STAR {Litton Kurisinkel, nfychen}@i2r.a-star.edu.sg Abstract We present set to ordered text, a natural lan- guage generation task applied to automatically generating discharge instructions from admis- sion ICD (International Classification of Dis- eases) codes. This task differs from other nat- ural language generation tasks in the following ways: (1) The input is a set of identifiable en- tities (ICD codes) where the relations between individual entities are not explicitly specified. (2) The output text is not a narrative descrip- tion (e.g. news articles) composed from the input. Rather, inferences are made from the input (ICD codes, which represent diagnoses and clinical procedures) to generate the output (instructions). (3) There is an optimal order in which each sentence (instruction) should ap- pear in the output. Unlike most other tasks, neither the input (ICD codes) nor their corre- sponding text representations of diagnoses and clinical procedures appear in the output, so the ordering of the output instructions needs to be learned in an unsupervised fashion. We hy- pothesize that each instruction in the output is mapped to a subset of ICD codes specified in the input. We propose a neural architecture that jointly models (a) subset selection: choos- ing relevant subsets from a set of input entities; (b) content ordering: learning the order of in- structions; (c) text generation: representing the instructions corresponding to the selected sub- sets in natural language. In addition, we penal- ize redundancy during beam search to improve tractability for long text generation. We for- mulate the problem setup and conducted ex- periments using the MIMIC-III dataset. Our model outperforms baseline models in both BLEU scores and human evaluations. 1 Introduction 1.1 Problem Statement Many healthcare applications exhibit a strong mapping between numerical or categorical infor- ICD1 ICD2 ICD3 ICD4 ICD5 ICD6 ICD7 ICD1 ICD3 ICD7 ICD1 ICD3 ICD7 ICD1 ICD3 ICD5 1. Instruction 1 2. Instruction 2 3. Instruction 3 Input ICD Set ICD Subsets Discharge Instructions Figure 1: Set to Ordered Text: Problem Definition mation and clinical instructions. To develop lan- guage generation capabilities for these settings, we define a task where discharge instructions are automatically generated using admission ICD (In- ternational Classification of Diseases) codes1 to potentially streamline clinical workflow. We de- fine a task where the input is a set of identifiable items (ICD codes) and the output consists of or- dered text sequences (instructions), which are in- ferred from the input (see Figure 1). 1.2 Proposed Approach We hypothesize that each discharge instruction in the output is mapped to a subset of ICD codes specified in the input. Our proposed approach thus models the correlations between individual enti- ties in the input set to choose the most …

Content truncated to 500 words